adding kernel name to the dispatch callback data

inc/rocprofiler.h

@@ -180,8 +180,9 @@ hsa_status_t rocprofiler_reset(rocprofiler_t* context,  // [in] profiling contex
 // Profiling callback data
 typedef struct {
   hsa_agent_t agent;
-  uint64_t kernel_object;
   uint64_t queue_index;
+  uint64_t kernel_object;
+  const char* kernel_name;
 } rocprofiler_callback_data_t;
 
 // Profiling callback type

src/core/intercept_queue.h

@@ -1,7 +1,9 @@
 #ifndef _SRC_CORE_INTERCEPT_QUEUE_H
 #define _SRC_CORE_INTERCEPT_QUEUE_H
 
+#include <amd_hsa_kernel_code.h>
 #include <dlfcn.h>
+
 #include <atomic>
 #include <iostream>
 #include <map>
@@ -90,8 +92,9 @@ class InterceptQueue {
         rocprofiler_group_t group = {};
         const hsa_kernel_dispatch_packet_t* dispatch_packet =
             reinterpret_cast<const hsa_kernel_dispatch_packet_t*>(packet);
-        rocprofiler_callback_data_t data = {obj->agent_info_->dev_id,
-                                            dispatch_packet->kernel_object, user_que_idx};
+        rocprofiler_callback_data_t data = {obj->agent_info_->dev_id, user_que_idx,
+                                            dispatch_packet->kernel_object,
+                                            GetKernelName(dispatch_packet)};
         hsa_status_t status = on_dispatch_cb_(&data, on_dispatch_cb_data_, &group);
         if (status == HSA_STATUS_SUCCESS) {
           Context* context = reinterpret_cast<Context*>(group.context);
@@ -145,6 +148,21 @@ class InterceptQueue {
     return (*header >> HSA_PACKET_HEADER_TYPE) & header_type_mask;
   }
 
+  static const char* GetKernelName(const hsa_kernel_dispatch_packet_t* dispatch_packet) {
+    const amd_kernel_code_t* kernel_code = NULL;
+    hsa_status_t status =
+        util::HsaRsrcFactory::Instance().LoaderApi()->hsa_ven_amd_loader_query_host_address(
+            reinterpret_cast<const void*>(dispatch_packet->kernel_object),
+            reinterpret_cast<const void**>(&kernel_code));
+    if (HSA_STATUS_SUCCESS != status) {
+      kernel_code = reinterpret_cast<amd_kernel_code_t*>(dispatch_packet->kernel_object);
+    }
+    amd_runtime_loader_debug_info_t* dbg_info = reinterpret_cast<amd_runtime_loader_debug_info_t*>(
+        kernel_code->runtime_loader_kernel_symbol);
+    const char* kernel_name = (dbg_info != NULL) ? dbg_info->kernel_name : NULL;
+    return (kernel_name != NULL) ? strdup(kernel_name) : NULL;
+  }
+
   static mutex_t mutex_;
   static const packet_word_t header_type_mask = (1ul << HSA_PACKET_HEADER_WIDTH_TYPE) - 1;
   static rocprofiler_callback_t on_dispatch_cb_;

src/util/hsa_rsrc_factory.cpp

@@ -83,8 +83,14 @@ HsaRsrcFactory::HsaRsrcFactory() {
   CHECK_STATUS("Error Calling hsa_iterate_agents", status);
 
   // Get AqlProfile API table
+  aqlprofile_api_ = {0};
   status = hsa_system_get_extension_table(HSA_EXTENSION_AMD_AQLPROFILE, 1, 0, &aqlprofile_api_);
   CHECK_STATUS("aqlprofile API table query failed", status);
+
+  // Get Loader API table
+  loader_api_ = {0};
+  status = hsa_system_get_extension_table(HSA_EXTENSION_AMD_LOADER, 1, 0, &loader_api_);
+  CHECK_STATUS("loader API table query failed", status);
 }
 
 // Destructor of the class

src/util/hsa_rsrc_factory.h

@@ -28,6 +28,7 @@ POSSIBILITY OF SUCH DAMAGE.
 #include <hsa.h>
 #include <hsa_ext_finalize.h>
 #include <hsa_ven_amd_aqlprofile.h>
+#include <hsa_ven_amd_loader.h>
 #include <stdint.h>
 #include <stdio.h>
 #include <stdlib.h>
@@ -223,6 +224,9 @@ class HsaRsrcFactory {
   // Return AqlProfile API table
   const hsa_ven_amd_aqlprofile_1_00_pfn_t* AqlProfileApi() const { return &aqlprofile_api_; }
 
+  // Return Loader API table
+  const hsa_ven_amd_loader_1_00_pfn_t* LoaderApi() const { return &loader_api_; }
+
  private:
   // System agents iterating callback
   static hsa_status_t GetHsaAgentsCallback(hsa_agent_t agent, void* data);
@@ -254,6 +258,9 @@ class HsaRsrcFactory {
 
   // AqlProfile API table
   hsa_ven_amd_aqlprofile_1_00_pfn_t aqlprofile_api_;
+
+  // Loader API table
+  hsa_ven_amd_loader_1_00_pfn_t loader_api_;
 };
 
 }  // namespace util

test/ctrl/tool.cpp

@@ -192,7 +192,9 @@ void dump_context(context_entry_t* entry) {
     const unsigned feature_count = entry->feature_count;
     FILE* file_handle = entry->file_handle;
 
-    fprintf(file_handle, "Dispatch[%u], kernel_object(0x%lx):\n", index, entry->data.kernel_object);
+    fprintf(file_handle,
+            "Dispatch[%u], queue_index(%lu), kernel_object(0x%lx), kernel_name(\"%s\"):\n", index,
+            entry->data.queue_index, entry->data.kernel_object, entry->data.kernel_name);
 
     status = rocprofiler_group_get_data(&group);
     check_status(status);